Using General Regression and Probabilistic Neural Networks To Predict Human Intestinal Absorption with Topological Descriptors Derived from Two-Dimensional Chemical Structures

Abstract

The objective of this study was to develop rapid and reliable methods to predict the percent human intestinal absorption (%HIA) of compounds based on their 2D descriptors. The analyzed data set included 86 drug and drug-like molecules and was the same as that studied by Wessel and co-workers. Instead of using three-dimensional descriptors such as polar surface area, which require lengthy computations, we employed only two-dimensional topological descriptors derived from information about the two-dimensional structure of molecules. The %HIA values were modeled using a general regression neural network (GRNN) and a probabilistic neural network (PNN), variants of normalized radial basis function networks. Both networks performed well to model the %HIA values. The root-mean square (rms) error was 22.8 %HIA unit for the external prediction set for a GRNN model, and 80% of the external prediction set was correctly classified for a PNN model, indicating the potential of our approach to estimate the %HIA values for a large set of compounds as virtual libraries.